T. Heskes. The use of being stubborn and introspective. In J. Dean, H. Cruse, and H. Ritter, editors, Proceedings of the ZiF Conference on Adaptive Behavior and Learning, pages 55-65, Bielefeld, 1994. 27  Home/Search   Document Details and Download   Summary   Related Articles   Check

....For strategies of how to select actions in a directed manner in this context, see [20, 21] and the references therein. 3.1.4 Other Related Approaches A di erent kind of queries is considered by Ratsaby [22] He allows queries for additional examples from a selected pattern class. Heskes [23] presents a learning algorithm for perceptron learning from an unreliable teacher that is applicable to learning in noisy and unstationary environments. He shows that the student can always outperform the teacher, if he she neglects some of the teacher s answers, particularly those the student is ....

T. Heskes. The use of being stubborn and introspective. In J. Dean, H. Cruse, and H. Ritter, editors, Proceedings of the ZiF Conference on Adaptive Behavior and Learning, pages 55-65, Bielefeld, 1994. 27

....one. However, contrasting to the Bayesian prescription, they do not preserve the ff Gamma1 behavior in the presence of output noise. Interestingly, the on line Bayesian algorithm also produces biological like side effects such as the well known blocking phenomena [21] indifference to outlyers [24], recency effect etc. The simple (unmodulated) Hebb rule does not present any of these idiosyncrasies (which, from a biological reverse engieneering viewpoint, may be the most informative phenomena) The modulation of the Hebb mechanism at the synaptic level due to behavioral factors (attention, ....

....learning algorithms, there exists a selection pressure for the development of modules which are needed to estimate the unknown but important quantities (type and level of noise, performance level, surprise level, etc. These on line estimators have already been developed for the perceptron [11, 22, 24]. This type of estimators are usual in the field of the so called robust statistics [23] The estimation procedure of the parameters which define the Bayesian algorithm corresponds to a second order learning process: the student not only must learn, but also learn how to learn. Conclusions and ....

Heskes T, The use of being stubborn and introspective, in Proceedings of the ZiF Conference on Adaptive Behavior and Learning (Bielefeld 1994).

....with norms denoted by B and J . In the presence of noise, instead of oe B , the student has access only to a corrupted version oe B . For example, for multiplicative noise, each teacher output is flipped independently with probability (Biehl et al. 1995, Copelli et al. 1996b, Copelli, 1997, Heskes, 1994): P ( oe B joe B ) 1 Gamma )ffi(oe B ; oe B ) ffi(oe B ; Gamma oe B ) 1) where oe B = sign(y) and y = B Delta S=B is the normalized field. The Kronecker ffi is 1 (0) only if the arguments are equal (different) In the same way, for the student, the field x = J Delta S=J and the ....

....noise does not alter fi for the optimized algorithms, additive (weight) noise does. This extension deserves a separate study. See (Biehl et al. 1995) for the behavior of the optimized algorithm and noise level estimation in the presence of noise acceleration in the absence of drift; see also (Heskes, 1994) where it is shown that learning is possible even in the mixed drift noise case. 5. The Wisconsin Test for Perceptrons: Piecewise Constant Rules How do the algorithms studied in the previous sections perform in the case of abrupt changes (piecewise constant rules) The interest is in determining ....

Heskes, T. (1994). The use of being stubborn and introspective, In J. Dean, H. Cruse & H. Ritter (Eds.) Proceedings of the ZiF Conference on Adaptive Behavior and Learning . University of Bielefeld, Bielefeld, Germany.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC