diff --git a/.github/workflows/python_tests.yml b/.github/workflows/python_tests.yml
index 3a5a3a2b..8b8df107 100644
--- a/.github/workflows/python_tests.yml
+++ b/.github/workflows/python_tests.yml
@@ -30,7 +30,7 @@ jobs:
     - uses: actions/cache@v2
       with:
         path: ${{ env.pythonLocation }}
-        key: ${{ env.pythonLocation }}-${{ hashFiles('setup.cfg') }}-${{ hashFiles('requirements-dev.txt') }}
+        key: ${{ env.pythonLocation }}-${{ hashFiles('setup.cfg') }}-${{ hashFiles('requirements-dev.txt') }}-version-1
 
     - name: Install dependencies
       run: |
diff --git a/docs/whats_new.rst b/docs/whats_new.rst
index ed0d7e2d..903d3637 100644
--- a/docs/whats_new.rst
+++ b/docs/whats_new.rst
@@ -82,7 +82,7 @@ Changelog
 - Channel types  are now available from a new ``ch_types_all`` attribute, and non-EEG channel names are now available from a new ``ch_names_non_eeg`` attribute from :class:`PrepPipeline <pyprep.PrepPipeline>`, by `Yorguin Mantilla`_ (:gh:`34`)
 - Renaming of ``ch_names`` attribute of :class:`PrepPipeline <pyprep.PrepPipeline>` to ``ch_names_all``, by `Yorguin Mantilla`_ (:gh:`34`)
 - It's now possible to pass ``'eeg'`` to ``ref_chs`` and ``reref_chs`` keywords to the ``prep_params`` parameter of :class:`PrepPipeline <pyprep.PrepPipeline>` to select only eeg channels for referencing, by `Yorguin Mantilla`_ (:gh:`34`)
-- :class:`PrepPipeline <pyprep.PrepPipeline>` will retain the non eeg channels through the ``raw`` attribute. The eeg-only and non-eeg parts will be in raw_eeg and raw_non_eeg respectively. See the ``raw`` attribute, by `Christian Oreilly`_ (:gh:`34`)
+- :class:`PrepPipeline <pyprep.PrepPipeline>` will retain the non eeg channels through the ``raw`` attribute. The eeg-only and non-eeg parts will be in raw_eeg and raw_non_eeg respectively. See the ``raw`` attribute, by `Christian O'Reilly`_ (:gh:`34`)
 - When a ransac call needs more memory than available, pyprep will now automatically switch to a slower but less memory-consuming version of ransac, by `Yorguin Mantilla`_ (:gh:`32`)
 - It's now possible to pass an empty list for the ``line_freqs`` param in :class:`PrepPipeline <pyprep.PrepPipeline>` to skip the line noise removal, by `Yorguin Mantilla`_ (:gh:`29`)
 - The three main classes :class:`~pyprep.PrepPipeline`, :class:`~pyprep.NoisyChannels`, and :class:`pyprep.Reference` now have a ``random_state`` parameter to set a seed that gets passed on to all their internal methods and class calls, by `Stefan Appelhoff`_ (:gh:`31`)
diff --git a/examples/ransac_comparison_autoreject.py b/examples/ransac_comparison_autoreject.py
new file mode 100644
index 00000000..b2cb1925
--- /dev/null
+++ b/examples/ransac_comparison_autoreject.py
@@ -0,0 +1,212 @@
+"""
+===============================================
+RANSAC comparison between pyprep and autoreject
+===============================================
+
+Next to the RANSAC implementation in ``pyprep``,
+there is another implementation that makes use of MNE-Python.
+That alternative RANSAC implementation can be found in the
+`"autoreject" package <https://github.com/autoreject/autoreject/>`_.
+
+In this example, we make a basic comparison between the two implementations.
+
+#. by running them on the same simulated data
+#. by running them on the same "real" data
+
+
+.. currentmodule:: pyprep
+"""
+
+# Authors: Yorguin Mantilla <yjmantilla@gmail.com>
+#
+# License: MIT
+
+# %%
+# First we import what we need for this example.
+import numpy as np
+import mne
+from scipy import signal as signal
+from time import perf_counter
+from autoreject import Ransac
+import pyprep.ransac as ransac_pyprep
+
+
+# %%
+# Now let's make some arbitrary MNE raw object for demonstration purposes.
+# We will think of good channels as sine waves and bad channels correlated with
+# each other as sawtooths. The RANSAC will be biased towards sines in its
+# prediction (they are the majority) so it will identify the sawtooths as bad.
+
+# Set a random seed to make this example reproducible
+rng = np.random.RandomState(435656)
+
+# start defining some key aspects for our simulated data
+sfreq = 1000.0
+montage = mne.channels.make_standard_montage("standard_1020")
+ch_names = montage.ch_names
+n_chans = len(ch_names)
+info = mne.create_info(ch_names=ch_names, sfreq=sfreq, ch_types=["eeg"] * n_chans)
+time = np.arange(0, 30, 1.0 / sfreq)  # 30 seconds of recording
+
+# randomly pick some "bad" channels (sawtooths)
+n_bad_chans = 3
+bad_channels = rng.choice(np.arange(n_chans), n_bad_chans, replace=False)
+bad_channels = [int(i) for i in bad_channels]
+bad_ch_names = [ch_names[i] for i in bad_channels]
+
+# The frequency components to use in the signal for good and bad channels
+freq_good = 20
+freq_bad = 20
+
+# Generate the data: sinewaves for "good", sawtooths for "bad" channels
+X = [
+    signal.sawtooth(2 * np.pi * freq_bad * time)
+    if i in bad_channels
+    else np.sin(2 * np.pi * freq_good * time)
+    for i in range(n_chans)
+]
+
+# Scale the signal amplitude and add noise.
+X = 2e-5 * np.array(X) + 1e-5 * np.random.random((n_chans, time.shape[0]))
+
+# Finally, put it all together as an mne "Raw" object.
+raw = mne.io.RawArray(X, info)
+raw.set_montage(montage, verbose=False)
+
+# Print, which channels are simulated as "bad"
+print(bad_ch_names)
+
+# %%
+# Configure RANSAC parameters
+n_samples = 50
+fraction_good = 0.25
+corr_thresh = 0.75
+fraction_bad = 0.4
+corr_window_secs = 5.0
+
+# %%
+# autoreject's RANSAC
+ransac_ar = Ransac(
+    picks=None,
+    n_resample=n_samples,
+    min_channels=fraction_good,
+    min_corr=corr_thresh,
+    unbroken_time=fraction_bad,
+    n_jobs=1,
+    random_state=rng,
+)
+epochs = mne.make_fixed_length_epochs(
+    raw,
+    duration=corr_window_secs,
+    preload=True,
+    reject_by_annotation=False,
+    verbose=None,
+)
+
+start_time = perf_counter()
+ransac_ar = ransac_ar.fit(epochs)
+print("--- %s seconds ---" % (perf_counter() - start_time))
+
+corr_ar = ransac_ar.corr_
+bad_by_ransac_ar = ransac_ar.bad_chs_
+
+# Check channels that go bad together by RANSAC
+print("autoreject bad chs:", bad_by_ransac_ar)
+assert set(bad_ch_names) == set(bad_by_ransac_ar)
+
+# %%
+# pyprep's RANSAC
+
+start_time = perf_counter()
+bad_by_ransac_pyprep, corr_pyprep = ransac_pyprep.find_bad_by_ransac(
+    data=raw._data.copy(),
+    sample_rate=raw.info["sfreq"],
+    complete_chn_labs=np.asarray(raw.info["ch_names"]),
+    chn_pos=raw._get_channel_positions(),
+    exclude=[],
+    n_samples=n_samples,
+    sample_prop=fraction_good,
+    corr_thresh=corr_thresh,
+    frac_bad=fraction_bad,
+    corr_window_secs=corr_window_secs,
+    channel_wise=False,
+    random_state=rng,
+)
+print("--- %s seconds ---" % (perf_counter() - start_time))
+
+# Check channels that go bad together by RANSAC
+print("pyprep bad chs:", bad_by_ransac_pyprep)
+assert set(bad_ch_names) == set(bad_by_ransac_pyprep)
+
+# %%
+# Now we test the algorithms on real EEG data.
+# Let's download some data for testing.
+data_paths = mne.datasets.eegbci.load_data(subject=4, runs=1, update_path=True)
+fname_test_file = data_paths[0]
+
+# %%
+# Load data and prepare it
+
+raw = mne.io.read_raw_edf(fname_test_file, preload=True)
+
+# The eegbci data has non-standard channel names. We need to rename them:
+mne.datasets.eegbci.standardize(raw)
+
+# Add a montage to the data
+montage_kind = "standard_1005"
+montage = mne.channels.make_standard_montage(montage_kind)
+raw.set_montage(montage)
+
+
+# %%
+# autoreject's RANSAC
+ransac_ar = Ransac(
+    picks=None,
+    n_resample=n_samples,
+    min_channels=fraction_good,
+    min_corr=corr_thresh,
+    unbroken_time=fraction_bad,
+    n_jobs=1,
+    random_state=rng,
+)
+epochs = mne.make_fixed_length_epochs(
+    raw,
+    duration=corr_window_secs,
+    preload=True,
+    reject_by_annotation=False,
+    verbose=None,
+)
+
+start_time = perf_counter()
+ransac_ar = ransac_ar.fit(epochs)
+print("--- %s seconds ---" % (perf_counter() - start_time))
+
+corr_ar = ransac_ar.corr_
+bad_by_ransac_ar = ransac_ar.bad_chs_
+
+# Check channels that go bad together by RANSAC
+print("autoreject bad chs:", bad_by_ransac_ar)
+
+
+# %%
+# pyprep's RANSAC
+
+start_time = perf_counter()
+bad_by_ransac_pyprep, corr_pyprep = ransac_pyprep.find_bad_by_ransac(
+    data=raw._data.copy(),
+    sample_rate=raw.info["sfreq"],
+    complete_chn_labs=np.asarray(raw.info["ch_names"]),
+    chn_pos=raw._get_channel_positions(),
+    exclude=[],
+    n_samples=n_samples,
+    sample_prop=fraction_good,
+    corr_thresh=corr_thresh,
+    frac_bad=fraction_bad,
+    corr_window_secs=corr_window_secs,
+    channel_wise=False,
+    random_state=rng,
+)
+print("--- %s seconds ---" % (perf_counter() - start_time))
+
+# Check channels that go bad together by RANSAC
+print("pyprep bad chs:", bad_by_ransac_pyprep)
diff --git a/examples/run_ransac.py b/examples/run_ransac.py
index 93f2705d..4a9b53f5 100644
--- a/examples/run_ransac.py
+++ b/examples/run_ransac.py
@@ -90,7 +90,7 @@
 # `raw` object. For more information, we can access attributes of the ``nd``
 # instance:
 
-# Check channels that go bad together by correlation (RANSAC)
+# Check channels that go bad together by RANSAC
 print(nd.bad_by_ransac)
 assert set(bad_ch_names) == set(nd.bad_by_ransac)
 
diff --git a/requirements-dev.txt b/requirements-dev.txt
index 7b9feef0..4737af7d 100644
--- a/requirements-dev.txt
+++ b/requirements-dev.txt
@@ -1,3 +1,4 @@
+git+git://github.com/autoreject/autoreject.git@master#egg=autoreject
 black
 check-manifest
 flake8