-
Token-based Decision Criteria Are Suboptimal in In-context Learning
Authors:
Hakaze Cho,
Yoshihiro Sakai,
Mariko Kato,
Kenshiro Tanaka,
Akira Ishii,
Naoya Inoue
Abstract:
In-Context Learning (ICL) typically utilizes classification criteria from probabilities of manually selected label tokens. However, we argue that such token-based classification criteria lead to suboptimal decision boundaries, despite delicate calibrations through translation and constrained rotation. To address this problem, we propose Hidden Calibration, which renounces token probabilities and u…
▽ More
In-Context Learning (ICL) typically utilizes classification criteria from probabilities of manually selected label tokens. However, we argue that such token-based classification criteria lead to suboptimal decision boundaries, despite delicate calibrations through translation and constrained rotation. To address this problem, we propose Hidden Calibration, which renounces token probabilities and uses the nearest centroid classifier on the LM's last hidden states. In detail, we use the nearest centroid classification on the hidden states, assigning the category of the nearest centroid previously observed from a few-shot calibration set to the test sample as the predicted label. Our experiments on 3 models and 10 classification datasets indicate that Hidden Calibration consistently outperforms current token-based calibrations by about 20%. Our further analysis demonstrates that Hidden Calibration finds better classification criteria with less inter-categories overlap, and LMs provide linearly separable intra-category clusters with the help of demonstrations, which supports Hidden Calibration and gives new insights into the conventional ICL.
△ Less
Submitted 24 June, 2024;
originally announced June 2024.
-
ns-3 Implementation of Sub-Terahertz and Millimeter Wave Drop-based NYU Channel Model (NYUSIM)
Authors:
Hitesh Poddar,
Tomoki Yoshimura,
Matteo Pagin,
Theodore S Rappaport,
Art Ishii,
Michele Zorzi
Abstract:
The next generation of wireless networks will use sub-THz frequencies alongside mmWave frequencies to enable multi-Gbps and low-latency applications. To enable different verticals and use cases, engineers must take a holistic approach to build, analyze, and study different parts of the network and the interplay among the lower and higher layers of the protocol stack. It is of paramount importance…
▽ More
The next generation of wireless networks will use sub-THz frequencies alongside mmWave frequencies to enable multi-Gbps and low-latency applications. To enable different verticals and use cases, engineers must take a holistic approach to build, analyze, and study different parts of the network and the interplay among the lower and higher layers of the protocol stack. It is of paramount importance to accurately characterize the radio propagation in diverse scenarios such as urban microcell (UMi), urban macrocell (UMa), rural macrocell (RMa), indoor hotspot (InH), and indoor factory (InF) for a wide range of frequencies. The 3GPP statistical channel model (SCM) is oversimplified and restricted to the frequency range of 0.5-100 GHz. Thus, to overcome these limitations, this paper presents a detailed implementation of the drop-based NYU channel model (NYUSIM) for the frequency range of 0.5-150 GHz for the UMi, UMa, RMa, InH, and InF scenarios. NYUSIM allows researchers to design and evaluate new algorithms and protocols for future sub-THz wireless networks in ns-3.
△ Less
Submitted 2 May, 2023;
originally announced May 2023.
-
Full-Stack End-To-End mmWave Simulations Using 3GPP and NYUSIM Channel Model in ns-3
Authors:
H. Poddar,
T. Yoshimura,
M. Pagin,
T. S. Rappaport,
A. Ishii,
M. Zorzi
Abstract:
Accurate channel modeling and simulation tools are vital for studying sub-THz and millimeter (mmWave) wideband communication system performance. To accurately design future high data rate, low latency wireless modems, the entire protocol stack must be appropriately modeled to understand how the physical layer impacts the end-to-end performance experienced by the end user. This paper presents a ful…
▽ More
Accurate channel modeling and simulation tools are vital for studying sub-THz and millimeter (mmWave) wideband communication system performance. To accurately design future high data rate, low latency wireless modems, the entire protocol stack must be appropriately modeled to understand how the physical layer impacts the end-to-end performance experienced by the end user. This paper presents a full stack end-to-end performance analysis in ns-3 using drop-based NYU channel model (NYUSIM) and 3GPP statistical channel model (SCM) in scenarios, namely urban microcell (UMi), urban macrocell (UMa), rural macrocell (RMa), and indoor hotspot (InH) at 28 GHz with 100 MHz bandwidth. Video data is transmitted at 50 Mbps using User Datagram Protocol (UDP), and we observe that the RMa channel is benign in non-line of sight (NLOS) for NYUSIM and 3GPP SCM as it exhibits no packet drops and yields maximum throughput (48.1 Mbps) and latency of $\sim$ 20 ms. In NLOS, for NYUSIM, the UMa and RMa channels are similar in terms of throughput and packet drops, and the latency in UMi and InH scenarios is 10 times and 25 times higher respectively compared to UMa. Our results indicate that mmWave bands can support data rates of 50 Mbps with negligible packet drops and latency below 150 ms in all scenarios using NYUSIM.
△ Less
Submitted 5 March, 2023; v1 submitted 23 February, 2023;
originally announced February 2023.
-
Rain-Code Fusion : Code-to-code ConvLSTM Forecasting Spatiotemporal Precipitation
Authors:
Takato Yasuno,
Akira Ishii,
Masazumi Amakata
Abstract:
Recently, flood damage has become a social problem owing to unexperienced weather conditions arising from climate change. An immediate response to heavy rain is important for the mitigation of economic losses and also for rapid recovery. Spatiotemporal precipitation forecasts may enhance the accuracy of dam inflow prediction, more than 6 hours forward for flood damage mitigation. However, the ordi…
▽ More
Recently, flood damage has become a social problem owing to unexperienced weather conditions arising from climate change. An immediate response to heavy rain is important for the mitigation of economic losses and also for rapid recovery. Spatiotemporal precipitation forecasts may enhance the accuracy of dam inflow prediction, more than 6 hours forward for flood damage mitigation. However, the ordinary ConvLSTM has the limitation of predictable range more than 3-timesteps in real-world precipitation forecasting owing to the irreducible bias between target prediction and ground-truth value. This paper proposes a rain-code approach for spatiotemporal precipitation code-to-code forecasting. We propose a novel rainy feature that represents a temporal rainy process using multi-frame fusion for the timestep reduction. We perform rain-code studies with various term ranges based on the standard ConvLSTM. We applied to a dam region within the Japanese rainy term hourly precipitation data, under 2006 to 2019 approximately 127 thousands hours, every year from May to October. We apply the radar analysis hourly data on the central broader region with an area of 136 x 148 km2 . Finally we have provided sensitivity studies between the rain-code size and hourly accuracy within the several forecasting range.
△ Less
Submitted 1 March, 2021; v1 submitted 30 September, 2020;
originally announced September 2020.
-
Generative Damage Learning for Concrete Aging Detection using Auto-flight Images
Authors:
Takato Yasuno,
Akira Ishii,
Junichiro Fujii,
Masazumi Amakata,
Yuta Takahashi
Abstract:
In order to monitor the state of large-scale infrastructures, image acquisition by autonomous flight drones is efficient for stable angle and high-quality images. Supervised learning requires a large data set consisting of images and annotation labels. It takes a long time to accumulate images, including identifying the damaged regions of interest (ROIs). In recent years, unsupervised deep learnin…
▽ More
In order to monitor the state of large-scale infrastructures, image acquisition by autonomous flight drones is efficient for stable angle and high-quality images. Supervised learning requires a large data set consisting of images and annotation labels. It takes a long time to accumulate images, including identifying the damaged regions of interest (ROIs). In recent years, unsupervised deep learning approaches such as generative adversarial networks (GANs) for anomaly detection algorithms have progressed. When a damaged image is a generator input, it tends to reverse from the damaged state to the healthy state generated image. Using the distance of distribution between the real damaged image and the generated reverse aging healthy state fake image, it is possible to detect the concrete damage automatically from unsupervised learning. This paper proposes an anomaly detection method using unpaired image-to-image translation map** from damaged images to reverse aging fakes that approximates healthy conditions. We apply our method to field studies, and we examine the usefulness of our method for health monitoring of concrete damage.
△ Less
Submitted 19 August, 2020; v1 submitted 26 June, 2020;
originally announced June 2020.
-
Trained Model Fusion for Object Detection using Gating Network
Authors:
Tetsuo Inoshita,
Yuichi Nakatani,
Katsuhiko Takahashi,
Asuka Ishii,
Gaku Nakano
Abstract:
The major approaches of transfer learning in computer vision have tried to adapt the source domain to the target domain one-to-one. However, this scenario is difficult to apply to real applications such as video surveillance systems. As those systems have many cameras installed at each location regarded as source domains, it is difficult to identify the proper source domain. In this paper, we intr…
▽ More
The major approaches of transfer learning in computer vision have tried to adapt the source domain to the target domain one-to-one. However, this scenario is difficult to apply to real applications such as video surveillance systems. As those systems have many cameras installed at each location regarded as source domains, it is difficult to identify the proper source domain. In this paper, we introduce a new transfer learning scenario that has various source domains and one target domain, assuming video surveillance system integration. Also, we propose a novel method for automatically producing a high accuracy model by fusing models trained at various source domains. In particular, we show how to apply a gating network to fuse source domains for object detection tasks, which is a new approach. We demonstrate the effectiveness of our method through experiments on traffic surveillance datasets.
△ Less
Submitted 2 March, 2020;
originally announced March 2020.
-
A study of trends in the effects of TV ratings and social media (Twitter) -- Case study 1
Authors:
Takuya Ueoka,
Akira Ishii,
Yasuko Kawahata
Abstract:
The Japanese TV program 'Drama A' is a drama broadcast from October to December 2016. The audience rating was sluggish, but this drama marked a high audience rating in 2016. Since it was popular from the middle, and it was speculated that there was a part related to social media in the popularity, we considered existing research methods as a case study. In this paper, we used a mathematical model…
▽ More
The Japanese TV program 'Drama A' is a drama broadcast from October to December 2016. The audience rating was sluggish, but this drama marked a high audience rating in 2016. Since it was popular from the middle, and it was speculated that there was a part related to social media in the popularity, we considered existing research methods as a case study. In this paper, we used a mathematical model of the hit phenomenon to examine the impact of audience assessment from social media from a sociophysical perspective. We got the same consideration as the audience rating per minute of video research. This paper is IEEE BIGDATA2018's Revised paper(Consideration on TV audience rating and influence of social media).
△ Less
Submitted 11 August, 2019;
originally announced September 2019.
-
Consensus formation Online using Sociophysics method
Authors:
Yasuko Kawahata,
Akira Ishii
Abstract:
Consensus formation and difference of opinion have long been the subject of research. However, relevant laws and systems within society are being updated to reflect the changes in information networks. Online environment has come to fulfill a major role as a real and concrete place of opposing opinions and consensus formation. In the future, quantitative findings on consensus formation, and findin…
▽ More
Consensus formation and difference of opinion have long been the subject of research. However, relevant laws and systems within society are being updated to reflect the changes in information networks. Online environment has come to fulfill a major role as a real and concrete place of opposing opinions and consensus formation. In the future, quantitative findings on consensus formation, and findings on relevant trends, must be summarized, and quantitative research related to trends likely to give rise to social and economic risk is required. Thus, the potential for comparing research related to consensus formation using actual data and an approach using a mathematical model was first investigated.
△ Less
Submitted 18 July, 2019;
originally announced July 2019.
-
The Influence of Social Media Writing on Online Search Behavior for Seasonal Events: The Sociophysics Approach
Authors:
Nozomi Okano,
Masaru Higashi,
Akira Ishii
Abstract:
Using seasonal topics as the study subject, in this study, we focus on the timing gap between social media writing and online search behavior. To conduct our analysis, we used the mathematical model of search behavior, comprising the sociophysics approach. The seasonal topics selected were St.Valentine's Day, Halloween and New Year countdown. We also picked up the event like Christmas and Hallowee…
▽ More
Using seasonal topics as the study subject, in this study, we focus on the timing gap between social media writing and online search behavior. To conduct our analysis, we used the mathematical model of search behavior, comprising the sociophysics approach. The seasonal topics selected were St.Valentine's Day, Halloween and New Year countdown. We also picked up the event like Christmas and Halloween. We analyzed the influence of blogs and Twitter on search behavior and found a deviation of interest in terms of timing. We also analyzed Japanese seasonal event of eating Eho-maki in February 3 and eels at the day of the ox in midsummer.
△ Less
Submitted 31 December, 2018;
originally announced January 2019.
-
Opinion Dynamics Theory for Analysis of Consensus Formation and Division of Opinion on the Internet
Authors:
Akira Ishii,
Yasuko Kawahata
Abstract:
The massive amount of text data on the web has facilitated research on the quantitative analysis of public opinion, which could not be visualized earlier. In this paper, we propose a new opinion dynamics theory. This theory that is intended to explain agreement formation and opinion breakup division in opinion exchanges on social media such as Twitter. With the popularization of the public network…
▽ More
The massive amount of text data on the web has facilitated research on the quantitative analysis of public opinion, which could not be visualized earlier. In this paper, we propose a new opinion dynamics theory. This theory that is intended to explain agreement formation and opinion breakup division in opinion exchanges on social media such as Twitter. With the popularization of the public network, we have become able to communicate with instantaneity and interactivity beyond the temporal and spatial constraints.Research on quantitatively analyzing the distribution of opinion on public opinion that has not been visualized so far utilizing massive web text data is progressing.Our model is based on the Bounded Confidence Model, that expresses opinions in as continuous quantity values. However, in the Bounded Confidence Model, it was assumed that people with different opinions move not in disregard but ignoring opinions. Furthermore, in our theory, it modeled so that it can expresser model incorporates the influence from of the external pressure outside and the phenomenon depending on the surrounding situation.
△ Less
Submitted 31 December, 2018;
originally announced December 2018.
-
Analysis of social media content and search behavior related to seasonal topics using the sociophysics approach
Authors:
Akira Ishii,
Toshimichi Wakabayashi,
Nozomi Okano,
Yasuko Kawahata
Abstract:
We studied the time interval between posting social media content and search action related to seasonal topics. The analysis was performed using a mathematical model of the search behavior as in the theory of sociophysics. As seasonal topics, the word cherry blossom was considered for spring, bikini for summer, autumn leaves for fall, and skiing for winter. We examined the influence of blogs and T…
▽ More
We studied the time interval between posting social media content and search action related to seasonal topics. The analysis was performed using a mathematical model of the search behavior as in the theory of sociophysics. As seasonal topics, the word cherry blossom was considered for spring, bikini for summer, autumn leaves for fall, and skiing for winter. We examined the influence of blogs and Twitter posts given the search behavior and found a time deviation of interest on these topics.
△ Less
Submitted 13 July, 2018;
originally announced July 2018.
-
Measurement of human activity using velocity GPS data obtained from mobile phones
Authors:
Yasuko Kawahata,
Takayuki Mizuno,
Akira Ishii
Abstract:
Human movement is used as an indicator of human activity in modern society. The velocity of moving humans is calculated based on position information obtained from mobile phones. The level of human activity, as recorded by velocity, varies throughout the day. Therefore, velocity can be used to identify the intervals of highest and lowest activity. More specifically, we obtained mobile-phone GPS da…
▽ More
Human movement is used as an indicator of human activity in modern society. The velocity of moving humans is calculated based on position information obtained from mobile phones. The level of human activity, as recorded by velocity, varies throughout the day. Therefore, velocity can be used to identify the intervals of highest and lowest activity. More specifically, we obtained mobile-phone GPS data from the people around Shibuya station in Tokyo, which has the highest population density in Japan. From these data, we observe that velocity tends to consistently increase with the changes in social activities. For example, during the earthquake in Kumamoto Prefecture in April 2016, the activity on that day was much lower than usual. In this research, we focus on natural disasters such as earthquakes owing to their significant effects on human activities in developed countries like Japan. In the event of a natural disaster in another developed country, considering the change in human behavior at the time of the disaster (e.g., the 2016 Kumamoto Great Earthquake) from the viewpoint of velocity allows us to improve our planning for mitigation measures. Thus, we analyze the changes in human activity through velocity calculations in Shibuya, Tokyo, and compare times of disasters with normal times.
△ Less
Submitted 13 June, 2017;
originally announced June 2017.
-
Position-sensitive propagation of information on social media using social physics approach
Authors:
Akira Ishii,
Takayuki Mizuno,
Yasuko Kawahata
Abstract:
The excitement and convergence of tweets on specific topics are well studied. However, by utilizing the position information of Tweet, it is also possible to analyze the position-sensitive tweet. In this research, we focus on bomb terrorist attacks and propose a method for separately analyzing the number of tweets at the place where the incident occurred, nearby, and far. We made measurements of p…
▽ More
The excitement and convergence of tweets on specific topics are well studied. However, by utilizing the position information of Tweet, it is also possible to analyze the position-sensitive tweet. In this research, we focus on bomb terrorist attacks and propose a method for separately analyzing the number of tweets at the place where the incident occurred, nearby, and far. We made measurements of position-sensitive tweets and suggested a theory to explain it. This theory is an extension of the mathematical model of the hit phenomenon.
△ Less
Submitted 2 June, 2017;
originally announced June 2017.
-
Mathematical model for hit phenomena and its application to analyze popularity of weekly tv drama
Authors:
Akira Ishii,
Akiko Kitao,
Tsukasa Usui,
Koki Uchiyama
Abstract:
Mathematical model for hit phenomena presented by A Ishii et al in 2012 has been extended to analyze and predict a lot of hit subject using social network system. The equation for each individual consumers is assumed and the equation of social response to each hit subject is derived as stochastic process of statistical physics. The advertisement effect is included as external force and the communi…
▽ More
Mathematical model for hit phenomena presented by A Ishii et al in 2012 has been extended to analyze and predict a lot of hit subject using social network system. The equation for each individual consumers is assumed and the equation of social response to each hit subject is derived as stochastic process of statistical physics. The advertisement effect is included as external force and the communication effects are included as two-body and three-body interaction. The applications of this model are demonstrated for analyzing population of weekly TV drama. Including both the realtime view data and the playback view data, we found that the indirect communication correlate strongly to the TV viewing rate data for recent Japanese 20 TV drama.
△ Less
Submitted 4 January, 2015;
originally announced January 2015.
-
Mathematical model for hit phenomena as stochastic process of interactions of human interactions
Authors:
Akira Ishii,
Hisashi Arakaki,
Naoya Matsuda,
Sanae Umemura,
Tamiko Urushidani,
Naoya Yamagata,
Narihiko Yoshda
Abstract:
Mathematical model for hit phenomena in entertainments in the society is presented as stochastic process of interactions of human dynamics. The model use only the time distribution of advertisement budget as input and the words of mouth (WOM) as posting in the social network system is used as the data to compare with the calculated results. The unit of time is daily. The WOM distribution in time i…
▽ More
Mathematical model for hit phenomena in entertainments in the society is presented as stochastic process of interactions of human dynamics. The model use only the time distribution of advertisement budget as input and the words of mouth (WOM) as posting in the social network system is used as the data to compare with the calculated results. The unit of time is daily. The WOM distribution in time is found to be very close to the residue distribution in time. The calculations for Japanese motion picture market due to the mathematical model agree very well with the actual residue distribution in time.
△ Less
Submitted 5 December, 2011;
originally announced December 2011.
-
Revenue Prediction of Local Event using Mathematical Model of Hit Phenomena
Authors:
Akira Ishii,
Takehiro Matsumoto,
Shinji Miki
Abstract:
Theoretical approach to investigate human-human interaction in society performed using a many-body theory including human-human interaction. The advertisement is treated as an external force. The word of mouth (WOM) effect is included as a two-body interaction between humans. The rumor effect is included as a three-body interaction between humans. The parameters to define the strength of human int…
▽ More
Theoretical approach to investigate human-human interaction in society performed using a many-body theory including human-human interaction. The advertisement is treated as an external force. The word of mouth (WOM) effect is included as a two-body interaction between humans. The rumor effect is included as a three-body interaction between humans. The parameters to define the strength of human interactions are assumed to be constant values. The calculated result explained well the two local events "Mizuki-Shigeru Road in Sakaiminato" and "the sculpture festival at Tottori" in Japan.
△ Less
Submitted 4 December, 2011;
originally announced December 2011.
