dekker.one
Group work for a Monash Research Methods course
Reports edits
+21
-20
+21
-20
mini_proj/report/waldo.tex
+21
-20
mini_proj/report/waldo.tex
···
302
302
Random Forest & 92.23\% & 0.92\\
303
303
\hline
304
304
\end{tabular}
305
-
\captionsetup{width=0.70\textwidth}
305
+
\captionsetup{width=0.80\textwidth}
306
306
\caption{Comparison of the accuracy and training time of each neural
307
307
network and traditional machine learning technique}
308
308
\label{tab:results}
309
309
\end{table}
310
310
311
-
We can see by the results that Deep Neural Networks outperform our benchmark
312
-
classification models, although the time required to train these networks is
313
-
significantly greater.
314
-
315
-
% models that learn relationships between pixels outperform those that don't
316
-
311
+
\par
312
+
We can see in these results that Deep Neural Networks outperform our benchmark classification models in terms of the accuracy they achieve.
313
+
However, the time required to train these networks is significantly greater.
314
+
An additional consideration is the extra layer of abstraction present in the FCN and not the CNN.
315
+
This may indicate that the FCN can achieve better accuracies, given more training time (epochs).
316
+
\\
317
+
% models that learn relationships between pixels outperform those that don't
318
+
\todo{
319
+
Discussion of the results:
320
+
- Was this what we expected to see?
321
+
- What was surprising?
322
+
- If you take learning time into account, are NN still as good?
323
+
- We also did say we would have these other measures, so we should at least try to include them. Then the question is also what do they show.
324
+
}
325
+
\par
317
326
Of the benchmark classifiers we see the best performance with Random
318
-
Forests and the worst performance with K Nearest Neighbours. As supported
319
-
by the rest of the results, this comes down to a models ability to learn
320
-
the hidden relationships between the pixels. This is made more apparent by
321
-
performance of the Neural Networks.
327
+
Forests and the worst performance with K Nearest Neighbours.
328
+
The low training time of the random forests could be due to the task being one of binary classification, and the traversal of binary trees being efficient resulting in low training time.
329
+
In terms of the models' accuracies, this is supported by the rest of the results and comes down to a model's ability to learn hidden relationships between pixels.
330
+
This is made more apparent by performance of the Neural Networks.
322
331
323
-
\section{Conclusion} \label{sec:conclusion}
332
+
\section{Conclusion} \label{sec:conclusion}
324
333
325
334
Image from the ``Where's Waldo?'' puzzle books are ideal images to test
326
335
image classification techniques. Their tendency for hidden objects and ``red
···
337
346
It would be interesting to investigate various of these methods further.
338
347
There might be quite a lot of ground that could be gained by using
339
348
specialized variants of these clustering algorithms.
340
-
341
-
342
-
Discussion of the results:
343
-
- Was this what we expected to see?
344
-
- What was surprising?
345
-
- If you take learning time into account, are NN still as good?
346
-
- We also did say we would have these other measures, so we should at least try to include them. Then the question is also what do they show.
347
-
348
349
\clearpage % Ensures that the references are on a separate page
349
350
\pagebreak
350
351
\bibliographystyle{alpha}