לוח שנה משולב

One of the fundamental challenges for describing intelligent systems is quantifying the balance between their physical - metabolic and energetic - requirements, and information processing requirements for sensing and acting. Both statistical mechanics and information theory provide many examples for such computational tradeoffs. The question is if we can extend these principles to living and intelligent systems that are far from thermodynamic equilibrium. Starting form Large Deviation Theory (the asymptotic theory behind statistical mechanics) we can obtain connections between costs and reward rates and control and sensing information rates, for systems in "metabolic information equilibrium" with stationary stochastic environments (Tishby & Polani, 2010). This result can be considered as the canonical equilibrium characterization for systems that obtain a certain value through interactions with a stochastic environment, but have no new learning (e.g. "stupid" cleaning robots). The affect of learning can be considered by revisiting the sub-extensivity of predictive information in stationary environments (Bialek, Nemenman & Tishby 2002) and combining it with the requirement of computational tractability of planning. We argue that planning is possible if the information flow terms remain proportional to the reward terms on the one hand, but still bounded by the sub-extensive predictive information on the other hand.
I will discuss the possible implications of this new computational principle to the emergence of hierarchical representations via a renormalization scheme for the Bellman equation - the canonical equation of planning and control.

: A longstanding question at the nexus of cognition and neuroscience concerns the distribution of the labor of learning across different brain systems: what are the different ways in which the brain learns? Recent research has focused on the role of the striatum and midbrain dopamine regions in habitual learning of stimulus-reward associations. However, emerging evidence suggests that the hippocampus – widely known for its role in building flexible memories – is also modulated by reward and innervated by dopamine. This raises new hypotheses about the role of the hippocampus in learning, the unique contributions of the hippocampus and the striatum, and the nature of the relationship between them. I will present studies that address these hypotheses using an integrative approach that combines functional imaging (fMRI) in healthy individuals with studies of learning in patients with selective damage to the striatum or the hippocampus. Converging data from these approaches suggests that both the striatum and the hippocampus contribute to learning, with distinct implications for how learned information guides decisions.

Mechanical forces typical of daily life are several billion times stronger than the force between two atoms, such as a carbon - carbon bond. Although light and heat are routinely used as conventional energy inputs to drive chemical reactions, harnessing mechanical energy for the same goal is not trivial. In the 1930s, Staudinger found that polymers are able to undergo mechanically driven chemical bond scission, and, today, we are on the verge of understanding and exploiting this process at an unprecedented level. In the beginning of the talk, the experimental basis for this interesting energy transformation process will be presented. Then, some new accounts of polymer mechanochemistry will be discussed in more detail: complete mechanochemical unzipping of polymers to monomers, followed by repolymerization; mechanochemical reactions induced by polymer swelling; and mechanochemical production of acid in a bulk polymer, a considerable advance towards self-healing applications.