_Schema d'architecture plus avancée (avec augmentation avant et après retrieval + travail d'intégration de l'extraction dans le processus de génération)_
</font>
*Survey on RAG Meeting LLM: Towards Retrieval-Augmented Large Language Models [Fan et al., 2023]* {.footer}
### Approche plus théorique de retriever et generateur
...
...
@@ -43,6 +40,8 @@ _Generateur comme fonction d'une paire (question, ensemble de documents) vers te
### Example
## Explainability in Artificial Intelligence
### Explainability in Artificial Intelligence
...
...
@@ -61,14 +60,14 @@ Aims:
:::: {.columns}
::: {.column width="50%"}
Interpretability
::: {.column width="40%"}
**Interpretability**
_How ?_
:::
::: {.column width="50%"}
::: {.column width="40%"}
Explanability
**Explanability**
_Why ?_
:::
...
...
@@ -83,53 +82,65 @@ _Why ?_
### MechIR [@parry_mechir_2025]
#### Mechanistic interpretability
Understand the internal mechanisms of neural networks by performing
causal interventions on specific model components
Understand the internal mechanisms of neural networks by **performing causal interventions** on specific model components
#### MechIR
- Encoder-only models
- For Information Retrieval
- For Information Retrieval models
- Identify components responsible for some behavior
- Activation Patching Technique
### Activation Patching [@chen_axiomatic_2024]
Let $\mathcal{D}$ be the set of documents and $\mathcal{Q}$ the set of questions
Let $Q \times D \subset \mathcal{Q}\times\mathcal{D}$ be a set of pairs of questions and documents
Let $Q \times \tilde{D}$ the same set of pairs but with perturbed documents
1. Forward pass all $Q\times D$
- record $o_{i,j}^e$ the output of each component $n_{i,j}, \forall e \in Q\times D$
- record $p_D$ the performance of the model
2. Forward pass all $Q\times \tilde{D}$
- record $o_{i,j}^\tilde{e}$ the output of each component $n_{i,j}, \forall \tilde{e} \in Q\times \tilde{D}$
- record $p_\tilde{D}$ the performance of the model
3. Rewrite $D, e, \tilde{D} \text{ and } \tilde{e}$ as
- $\hat{D}, \hat{e}, \check{D} \text{ and } \check{e}$ if $p_D > p_\tilde{D}$
- $\check{D}, \check{e}, \hat{D} \text{ and } \hat{e}$ otherwise
4. For each component $n_{i,j}$ forward pass $Q\times\check{D}$ but replace $o_{i,j}^{\check{e}}$ by $o_{i,j}^{\hat{e}}$ for each $\check{e}$. Record the performance $\bar{p}$
5. $P = \frac{\bar{p} - p_\hat{D} }{\p_\check{D} - p_\hat{D}}$ gives the impact of the perturbation on the model performance
_Obtention du résultat + sens associé_
### Animation de l'execution de Activation patching
### Perturbation
Fonction qui pour chaque document, effectue la même modification.
Function that applies the same modification on each document.
Example :
``` {python}
#| echo: true
def perturbation(doc):
return doc.replace("microwave", "toaster")
```
Ici 1 type nous intéresse : remplacer un mot par un autre
#### Perturbation creation technique
- Identify vocabulary specific to the dataset
- find in the vocabulary words with several meaning $m_D$ and $m_D$
- Find in the vocabulary words with several meaning $m_D$ and $m_D$
- Replace that word by a synonym of the $m_D$ meaning
