Begin with a new installation of MFA, and head to the directory. First run validate with the original dictionary.

mfa validate ~/Downloads/tb/my_corpus english_us_arpa english_us_arpa

We see that there is in deed an section of corpus that is out-of-vocab.

INFO -     11 OOV word types
INFO -     18 total OOV tokens

Therefore, we will generate a new dictionary based on the existing dictionary of english_us_arpa.

First download the english_us_arpa model

mfa model download g2p english_us_arpa

Then, perform the actual dictionary generation:

• Want to interview more severe ashma
• Want to find someone younger

Difference between marketing and purchaser.

Taking to people

Spoke with Matt. Talked with more details with prototyping and how they can build a unique product.

Have not gotten back to him yet.

Recall, from conditional plan evaluation, we had that:

\begin{equation} U^{\pi}(b) = \sum_{s}^{} b(s) U^{\pi}(s) \end{equation}

let’s write it as:

\begin{equation} U^{\pi}(b) = \sum_{s}^{} b(s) U^{\pi}(s) = {\alpha_{\pi}}^{\top} b \end{equation}

where \(\U_{\pi}(s)\) is the conditional plan evaluation starting at each of the initial states.

\begin{equation} \alpha_{\pi} = \qty[ U^{\pi}(s_1), U^{\pi}(s_2) ] \end{equation}

You will notice, then the utility of \(b\) is linear on \(b\) for different policies \(\alpha_{\pi}\):

At every belief \(b\), there is a policy which has the highest \(U(b)\) at that \(b\) given be the alpha vector formulation.

see also in programming string (C)

• an alphabet \(\Sigma\) is a finite set
• a finite-sequence of elements in \(\Sigma\) is called a string
• the set of all strings in \(\Sigma\) is called \(\Sigma^{*}\), which includes the empty string
• for a particular string \(x\), the length of it is \(|x|\)
• the string of length zero is called \(\varepsilon\)
• a language is a subset of \(\Sigma^{*}\), meaning its a set of strings

Omer seems to call strings “words” sometimes.

languages are boolean function over strings

For every language \(L\) over \(\Sigma\) corresponds to a unique function \(f: \Sigma^{*} \to \{0,1\}\), whereby if \(f(x) = 1\), then \(x \in L\); otherwise, if \(f(x) = 0\), \(x \not \in L\).

Alternating Least Squares is a method to Factoring a matrix into two components:

\begin{equation} \mathcal{M}( R) \approx \mathcal{M}(U) \cdot \mathcal{M}(P) \end{equation}

where, we want to come up matricies \(U\) and \(P\) with a certain side length \(k\) that we exdogenously come up with

To perform Alternating Least Squares, we fix the values of either \(U\) or \(P\), then perform the least-squares optimization on

(This is proven best-fit for “non-pathological matricies”)