Population Codingの 最近の話題から 2018/12/3 Population Codingの 最近の話題から 銅谷賢治 doya@atr.co.jp ATR 人間情報科学研究所 科学技術振興事業団 CREST
あらすじ 古典的population coding仮説 Population vector 確率的population coding仮説 Bayes/最尤推定 不確定性の表現 複数入力の統合
Population Vectors (Georgopoulos) 1次運動野ニューロンの運動方向選択性 cosine tuning fi(m) = bi + ai cos( qm-qi) preferred direction qi は一様に分布 Population vector v(m) = Si fi(m) vi simplicity and robustness.
Probabilistic Population Codes (Zemel, Dayan, Pouget 1998) Encoding: underlying quantity x noisy response: P[r|x] Decoding Bayesian: P[x|r] P[r|x] P[x] Standard Poisson model encode single value of x P[ri|x] = e-fi(x) (fi(x))ri/ri! P[x|r] = P[x] Pi e-fi(x) (fi(x))ri/ri!
Extended Poisson Model Encode the probability distribution P[x|w] uncertainty, multiple values <ri> = x P[x|w] fi(x) dx Decoding for histogram representation <ri> = Sj fj fij logP[{fj}|{ri}] = K + logP[{fj}] + Si ri log[Sj fj fij]
Roles of the Cerebral Cortex Representations for good generalization modality, invariance, resolution,... task oriented context and working memory Combining multiple outputs reasons for for poplulation coding?
Population Coding (Deneve et al., 2001) Function approximation and Cue Integration
おわりに 情報の表現から処理へ Poisson的な発火の積極的な意味付けは?