Echo State Networks and Reservoir Computing Quiz

Answers and explanations appear as you take the quiz.

1. In an Echo State Network, which weights are actually trained?
   • Only the output (readout) weights
   • All recurrent weights inside the reservoir
   • Only the input weights
   • No weights are ever trained
2. Which condition expresses the echo state property for two runs with the same input history?
   • Their recurrent weights must all be trained by backpropagation
   • Their reservoir states eventually become independent of different initial states
   • Their output weights must have spectral radius exactly one
   • Their hidden states must equal the raw input at every step
3. Why is training an ESN much faster than training a standard LSTM?
   • ESNs use no neurons at all
   • ESNs run only on quantum computers
   • ESNs avoid backpropagation-through-time by only solving a linear readout
   • ESNs require no input data
4. How is spectral radius commonly used when tuning an echo-state reservoir?
   • It directly sets the number of readout classes
   • It guarantees fading memory for every recurrent matrix
   • It measures the number of trainable reservoir weights
   • It is a recurrent-feedback scale heuristic for balancing rich dynamics and stability, not a complete guarantee of the echo state property
5. Liquid State Machines, a spiking-neuron form of reservoir computing, were introduced by whom?
   • Wolfgang Maass
   • Geoffrey Hinton
   • Andrew Ng
   • Yoshua Bengio
6. What does the fixed reservoir compute before the ESN readout produces an output?
   • A normalized probability over all possible weights
   • A high-dimensional dynamic state representation of recent inputs
   • A learned transition table for every future sequence
   • A separate trained policy for each output class
7. What method is commonly used to fit an Echo State Network’s readout weights?
   • Backpropagation through all fixed recurrent weights
   • Dynamic routing by agreement
   • Linear least-squares regression, often with ridge regularization
   • A softmax over reservoir eigenvalues
8. What does the spectral radius of a recurrent weight matrix mean?
   • The sum of all matrix entries
   • The number of nonzero output weights
   • The average magnitude of the input sequence
   • The largest absolute value among its eigenvalues