Title: Cognitively motivated deep neural representations and architectures
Abstract: Despite tremendous progress in artificial intelligence we continue to lag behind human capabilities in crucial areas such as generalization, robustness and efficiency - hallmarks of human cognition. This talk proposes a paradigm shift towards
cognitively-motivated representations that explicitly incorporate macroscopic cognitive principles such as low-dimensionality, hierarchy, abstraction, neural feedback and sparsity. First, we argue that traditional metric spaces and linear tools are poorly
suited for efficient information storage and processing, contrasting sharply with the brain's more effective organizational strategies. We show that a top-down hierarchical manifold representation of low-dimensional, sparse subspaces can achieve human-like
performance in both decoding and induction tasks, particularly in lexical semantics. Next we explore the role of feedback mechanisms in the brain, especially feedback-driven deactivation of cortical columns, and present our work on MMLatch, a bottom-up top-down
fusion model applied to multimodal sentiment analysis. This research highlights the importance of bidirectional information flow in neural architectures. Finally, we discuss a novel neural network architecture inspired by synaptic pruning during brain development.
This approach utilizes long connections instead of traditional short residual connections, naturally pushing information to the first few layers of the network and resulting in sparsity. These networks exhibit behaviors reminiscent of biological brain networks,
including enhanced robustness to noise, good performance in low-data settings, and longer training times. Overall, our work demonstrates that by embracing cognitively-motivated principles in AI architectural design, we can create more efficient, robust, and
human-like AI systems capable of improved generalization and induction.
Bio: Alexandros Potamianos received the Diploma in electrical and computer engineering from the National Technical University of Athens, Greece, in 1990, and the M.S. and Ph.D. degrees in engineering sciences from Harvard University, Cambridge, MA, in
1991 and 1995, respectively. From 1995 to 1999, he was a Senior Technical Staff Member with AT&T Shannon Labs, Florham Park, NJ. From 1999 to 2002, he was a Technical Staff Member and Technical Supervisor with Bell Labs, Lucent Technologies, Murray Hill, NJ.
From 2003 to 2013, he served as an associate professor at the Department of ECE, Technical University of Crete, Chania, Greece. Since 2013, he serves as an associate professor at the School of ECE, National Technical University of Athens, Greece. He is also
a visiting professor at the Viterbi School of Engineering, University of Southern California, CA and an Amazon Scholar. He is the co-founder of Behavioral Signals, an emotion AI deep tech startup. He has authored or coauthored over 200 papers in professional
journals and conferences, and holds five patents. His current research interests include foundation models. speech processing, dialog and multimodal systems, natural language understanding, machine learning and multimodal child-computer interaction. Prof.
Potamianos has served multiple terms at the IEEE Speech and Language Technical Committee and at the IEEE Multimedia Technical Committee. He received a 2005 IEEE Signal Processing Society Best Paper Award. He is an IEEE fellow, an International Speech Communication
Association (ISCA) fellow and a fellow of the
Asia-Pacific Artificial Intelligence Association (AAIA).
________________________________________________________________________________
Meeting ID:
338 882 842 320
Passcode:
Vsw5aV
________________________________________________________________________________