|
Research /
PapersReinforcement Learning
Thesis on using options in TD-networks (prediction networks). Prediction is a useful abstraction, and allows us to predict our own predictions recursively (and therefore arbitrarily distant into the future). Great review of basic RL and history / prediction based learning frameworks.
Schmidhuber’s idea of an RL system that can modify its own Turing code (both the proof searching part, and the utility function computing aspects, etc.) by proving that improvements are useful first, and then implementing them.
The authors argue that representations formed by schema-learning can be turned into (or may actually be the equivalent of) predistive state representations.
Rafols provides a thorough review of the RL literature on representation, then describes a novel method of temporally abstracting options.
AINaur argues that human thinking is not the equivalent of computing, elaborates on the key distinctions, provides his own association-network model of human thinking (which he implies has been muzzled by academia), and suggests that psychology has basically gone downhill since William James. His points on the lack of grounded description and empirical observation in modern psychology are especially poignant.
Addiction, Reward, and Affect:
Robotics
Gates announces that the era of consumer robotics is upon us, and that microsoft is investing significant amounts of money into development tools (Microsoft Robotics Studio) to make it happen.
Philosophy of KnowledgeNeural Networks
When using the Least-Mean-Squared (LMS) technique (and thereby also backpropagation), “a neural network can be considered a nonparametric technique for estimation of a-posteriori probabilities.” (see end of Section IV on pg. 2) PDF
Computational Neuroscience:
“We have reviewed evidence that supports the proposal that dopamine neurons in the VTA and the substantia nigra report ongoing prediction errors for reward. The output of these neurons is consistent with a scalar prediction error signal; therefore, the delivery of this signal to target structures may influence the processing of predictions and the choice of reward-maximizing actions.”
“cortical mechanisms interact with hippocampal time dilation and contraction, amygdala “toggling” of salient features, and striatal reinforcement learning in cases of relevant feedback. Together the system produces incrementally constructed and selectively reinforced hierarchical representations consisting of nested sequences of clusters (Granger 2006).”
NeuroscienceEric Kandel’s autobiography, published following his nobel prize award in physiology for his contributions to the neural mechanisms of learning. Most notably, Kandel and his colleagues studied LTP and LTD through habituation, sensitization, and classical conditioning in the sea slug Aplysia.
Your genetics affects how likely your body is to amplify the effects of brain trauma and alzheimers.
“Both types of attachment activated regions specific to each, as well as overlapping regions in the brain’s reward system that coincide with areas rich in oxytocin and vasopressin receptors. Both deactivated a common set of regions associated with negative emotions, social judgment and ‘mentalizing’, that is, the assessment of other people’s intentions and emotions. We conclude that human attachment employs a push– pull mechanism that overcomes social distance by deactivating networks used for critical social assessment and negative emotions, while it bonds individuals through the involvement of the reward circuitry, explaining the power of love to motivate and exhilarate.”
Various anatomical images and fMRI scans of diseases, etc.
Psychology |