The long title translates as A theoretical approach to connectionist learning, with applications for speech recognition. This is a relatively big document with a number of relatively new ideas for the time.

• Connectionist learning algorithms can be studied as stochastic approximations and more specifically stochastic gradient descent algorithms, possibly involving surrogate loss functions (chapters 2 and 3).
• The performance of these algorithms can be studied using advanced statistical theories such as Vladimir Vapnik's Structural Risk Minimization (chapter 4).
• The speed of these algorithms can be approached using the methods of numerical optimization (chapter 5.)
• Complex applications, such as speech recognition, can be addressed using modular learning systems such as Discriminant Hidden Markov Models (chapter 8 and 9).
• Probabilistic discriminant modular learning systems are often limited by the so-called label-bias problem (chapter 10). Finding a solution took a three more years.

@phdthesis{bottou-91a,
  title = {Une Approche th\'eorique de l'Apprentissage Connexionniste: Applications \`a la Reconnaissance de la Parole},
  author = {Bottou, {L\'eon}},
  year = {1991},
  school = {Universit\'{e} de Paris XI},
  address = {Orsay, France},
  url = {http://leon.bottou.org/papers/bottou-91a},
}